MacTrack2’s Input Folders Examples

Introduction

As you may have seen if you already looked up the two input folders (input_model and input_tracking), they are almost identical — because they are copies. We decided to provide you with two separate folders to help you with the two main processes supported by this project:

  • Model building, training, and testing

  • Tracking: Segmentation and analysis of a video using a model

You can use them as examples. The module we primarily use in this project, kartezio, is quite restrictive regarding input formats. Hence, we aim to make it easier for you to understand.

Input Folders

Tracking

The structure of the input_tracking folder is as follows:

input_tracking/
  ├── dataset
  ├── models
  ├── vert
  │   ├── frames  # (empty)
  │   └── greenchannelvideo.avi
  └── redchannelvideo.avi

Model Building

The structure of the input_model folder is as follows:

input_model/
  ├── dataset
  ├── models

There are no videos here, as they are not necessary for model creation.

Contents

Dataset

The dataset folder is organized as follows:

dataset
├── test
│   ├── test_x
│   └── test_y
├── train
│   ├── train_x
│   └── train_y
├── dataset.csv
└── META.json

In order for kartezio to function correctly, this structure is required. Without it, errors will occur.

We will now describe each subfolder:

  • The train folder contains the images selected for model training. For instance, in the provided example, there are 25 microscopic images in train_x.

  • The train_y folder contains corresponding ground truth masks, created manually using ImageJ (see the Materials and Methods section of the README).

  • This training folder will remain the same in both input_model and input_tracking, as it is used for structure during tracking.

The test folder will change depending on your use case:

  • For testing a model (i.e., in the input_model folder), segment a few images and store them just as with the training set.

  • For tracking and video segmentation, the frames extracted from your video will replace the test images. This is a limitation of kartezio.

That is why we provide both folders and recommend creating your own by copying them. Be sure to store and track your model folders carefully.

  • dataset.csv and META.json are also necessary for kartezio to properly interpret the input folder.

If you want to create your own model — by retraining the provided one or building from a new dataset — follow the steps in quickstart.py.

Models

The models folder contains a hash-named directory with two JSON files:

  • elite.json: Contains the final pipeline of the model

  • history.json: Contains the training history (e.g., generations)

Specificities for Tracking

In the input_tracking folder, you will find:

  • A red channel video

  • A vert folder containing: - A green channel video - An empty frames folder

These are example videos for tracking. If you wish to segment your own video, you must respect this structure.

Both videos (red and green channels) originate from the same microscopic analysis, with channels separated using ImageJ. For more information, refer to the Materials and Methods section in the README.